walkthrough

Files are grouped by the behavior they serve. For each part of the change, you see the decision, decisive evidence, verification, and remaining risk, with a concrete example. You decide whether to accept each part before the walkthrough continues.

Install the skill with the command above. Invoke it explicitly: /walkthrough in Claude or $walkthrough in Codex. Name the change, for example /walkthrough this branch.

Use it to understand the last completed change, working tree, branch, or PR before approval. For a written list of findings, use review.

The agent verifies the selected change and reports working-tree changes outside it. It builds an ordered map accounting for every changed file, then presents one part at a time.

For the main behavior, it offers a small exercise, such as running a command or opening a page, and asks what you expect to see. You may skip the exercise. Each part ends with one acceptance question and waits for your answer.

Corrections are recorded without being implemented. A fix needs separate authorization. After a fix changes the code, the agent revisits the affected parts. The walkthrough ends when every changed file is covered and every correction is resolved or explicitly deferred.

You can request a separate read-only presenter agent. Your agent checks the evidence that presenter supplies.

SKILL.md links the presentation patterns and optional log format.

• The behavior and files covered, with your decisions and corrections.
• Verification evidence, its limits, and unresolved risks or questions.
• A walkthrough log, if requested.

Silence, a passing test, or completing an exercise does not count as approval. Apart from an authorized log, walkthrough changes no files and does not commit, push, or publish.